【Frontiers in HaiChuan Chemical Technology】Peking University team successfully solves the challenge of electrolyte design

Thread Content

The 25th National \"Safety Production Month\" in 2026: Everyone talks about safety, and everyone knows how to handle emergencies; identifying and addressing risks and hazards. -------------------------------------------------- Recently, a team led by Professor Pang Quanquan from the School of Materials Science and Engineering at Peking University, in collaboration with top domestic and international research institutions and enterprises such as Tsinghua University, Lawrence Berkeley National Laboratory, Princeton University, and SESAI Corp., has made breakthrough advances in the development of electrolytes for lithium-metal batteries. Through team innovation, a two-stage research framework that integrates deep active learning with knowledge transfer was proposed, successfully addressing the industry challenges associated with electrolyte design – namely the difficulties in trial and error, long development cycles, and high costs – thereby enabling the rapid identification of high-performance electrolytes and efficient transfer of relevant knowledge. The relevant research findings were published online in advance on March 27 in the top international journal Nature Communications, providing key technical support for the commercialization of next-generation high-energy-density lithium-metal batteries. Lithium metal batteries are recognized as the key development direction for next-generation energy storage and power batteries due to their extremely high theoretical energy density, and advancements in their performance are of great significance for the upgrading of the new energy industry. However, core challenges such as the low Coulombic efficiency of lithium metal anodes and poor interfacial stability have long restricted their large-scale application. As a key component that regulates the anode interface and determines the battery’s cycle life, the electrolyte is crucial for overcoming these bottlenecks. Unlike the development of traditional materials, electrolyte design is confronted with a vast and discrete chemical search space; lithium salts, solvents, additives, and various concentration combinations give rise to an enormous number of formulations. Moreover, the interfacial chemistry between components and the chemistry of the electrolyte are interconnected, resulting in highly discontinuous relationships between performance parameters. The traditional \"trial-and-error\" approach to R&D not only incurs high experimental costs and long development cycles, but also struggles to handle complex scenarios such as the introduction of new molecules and the expansion of high-dimensional formulations; therefore, the industry urgently needs a new, efficient, and intelligent paradigm for electrolyte development. To address the three key challenges in electrolyte design – a large search space, discontinuous performance relationships, and high experimental noise – Pang Quanquan’s team, in collaboration with researchers from various disciplines and institutions, developed an innovative two-stage framework that combines deep active learning (DAL) with target statistical coding (TSC). By leveraging artificial intelligence technologies, this framework enables intelligent screening of electrolytes and the transfer of design knowledge, allowing for the precise identification of high-performance formulations even with limited experimental data, as well as the efficient application of the learned design principles to complex scenarios. In the first phase of the research, the team focused on a space of 720 initial electrolyte formulations composed of lithium salts, solvents, additives, and concentrations. They innovatively employed deep kernel learning combined with the Thompson sampling algorithm to intelligently select the most informative experimental samples in each iteration. This approach enabled them to accurately model the highly non-linear and discontinuous relationship between electrolyte formulations and battery cycle life, thereby avoiding ineffective experiments and significantly improving the efficiency of the selection process. In the second phase, the team uses target statistical coding techniques to explicitly encode the complex correlations between components identified during the active learning process, thereby creating a reusable and transferable knowledge framework for electrolyte design. This approach overcomes the limitations of relying on single-formula approaches, enabling the transfer of such knowledge to higher-dimensional candidate spaces, as well as to lithium-metal full batteries and scenarios involving new molecular formulations. The experimental results fully demonstrate the efficiency and accuracy of this framework: within a space of 720 initial formulations, after just three rounds of deep active learning iterations and a total of 128 battery sample tests, the average cycle life of the batteries increased from 41.9 cycles in the randomly selected phase to 125.1 cycles ； The share of short-life batteries dropped sharply from 80.6% to 28.1%, while the share of long-life batteries increased from 9.7% to 40.6%. The top 5 high-quality electrolytes selected were subjected to repeated verification, and their overall performance was significantly superior to that of the high-performance formulations reported in existing literature of the same type, demonstrating the reliability and advancement of the screening method. More importantly, the research achieved efficient cross-scenario transfer of electrolyte design knowledge, overcoming the challenges in developing complex systems. By expanding the initial 720 formulations to a higher-dimensional candidate space of 5,400 options, high-quality electrolytes can be identified rapidly without any sample data required. The average cycle life of the top 5 formulations reached 200.6 cycles, which is 1.6 times higher than the best results obtained in the original space ； In a lithium metal/NCM811 full-cell system close to practical applications, the average capacity retention after 100 cycles of electrolyte migration was 84.0%, which is far higher than the 58.2% achieved with the initial formulation ； Faced with the combinatorial explosion caused by the introduction of new molecules, a space of 5,760 new formulations was created. Through just one round of experiments using 32 samples, the average capacity retention rate after 150 cycles increased from 24.4% to 56.5%; the optimal formulation retained 83% of its capacity even after 250 cycles, meeting the requirements for the development of new electrolytes. For the first time, this study integrates deep active learning with knowledge transfer in a comprehensive manner, offering a new intelligent research and development paradigm for complex electrolyte systems that features \"few samples, high efficiency, and transferability.\" It overcomes the limitations of traditional trial-and-error approaches to research and development, and opens up new technical pathways for addressing the \"carbonate/ether conflict\" in lithium-metal batteries as well as for developing long-lasting, high-performance electrolytes. Research findings clearly demonstrate that artificial intelligence-driven autonomous material discovery is facilitating a shift in the development of new energy materials from an experience-based approach to a data-driven one, and from inefficient trial and error to precise prediction, thereby providing important support for accelerating the industrialization of lithium-metal batteries. This research paper was jointly completed by Peking University, Tsinghua University, Lawrence Berkeley National Laboratory, Princeton University, and SESAICorp. Professor Pang Quanquan from Peking University, Associate Professor Jiang Benben from Tsinghua University, and Xu Kang from SESAICorp. serve as co-corresponding authors of the paper. Hong Xufeng, a 2025 doctoral graduate from Peking University, is the first author, while Wang Xizhe, a doctoral student at Tsinghua University, is the co-first author. Peking University is the primary institution affiliated with this paper. The research work has received support from various funding programs, including the **Key R&D Program**, the **Natural Science Foundation**, the Tsinghua-Toyota Joint Research Fund, the Beijing Natural Science Foundation, the Beijing Information Science and Technology **Research Center**, as well as 111 International Cooperation Projects.